A Single Malicious Webpage Let Researchers Hijack Meta Muse and Access Private Chats
The most striking result was silent access: a malicious page could send a user's Muse conversations and stored memories to an external server without an approval prompt. It could also alter those records. That exposed private information held by the assistant, not just the content of the page being viewed. In NebuSec's laboratory demonstration, Muse opened a malicious page. The researchers used it to extract conversation history and stored memories, then change them. They chained a flaw in Muse's built-in browser to a Linux kernel flaw, taking full control of the user's cloud computer. From there, they reached files and security components normally walled off from the agent. The test was not an attack on real users, and it stayed on one user's cloud machine. NebuSec says Muse's team fixed all the weaknesses before publication. Meta has not publicly confirmed the fixes, and the write-up does not say whether anyone's data was accessed before patching. The demonstration shows a possible exposure, not evidence that users were compromised.
What did the malicious webpage let researchers do to a Muse user's conversations and stored memories?
The most striking result was silent access: a malicious page could send a user's Muse conversations and stored memories to an external server without an approval prompt. It could also alter those records. That exposed private information held by the assistant, not just the content of the page being viewed.
In NebuSec's laboratory demonstration, Muse opened a malicious page. The researchers used it to extract conversation history and stored memories, then change them. They chained a flaw in Muse's built-in browser to a Linux kernel flaw, taking full control of the user's cloud computer. From there, they reached files and security components normally walled off from the agent.
The test was not an attack on real users, and it stayed on one user's cloud machine. NebuSec says Muse's team fixed all the weaknesses before publication. Meta has not publicly confirmed the fixes, and the write-up does not say whether anyone's data was accessed before patching. The demonstration shows a possible exposure, not evidence that users were compromised.
How far did the demonstrated attack reach, and what systems did it not reach?
The demonstrated attack reached the individual cloud computer running one user's Muse agent. Once the researchers had full control of that machine, they could reach saved files, the credential store, and Sentinel. These were parts of Muse's security design intended to be kept beyond the agent's reach.
The researchers began with a malicious page that could extract and alter conversation history and memories. They then chained a flaw in Muse's built-in browser to a Linux kernel flaw. That sequence gave them control of the user's cloud computer, extending the compromise beyond the assistant's conversation data and into protected parts of its environment.
The boundary held at the machine level: the researchers did not break out to the shared server beneath it or into other users' machines. The work was a laboratory demonstration, not an attack on real users. NebuSec says the weaknesses were fixed before publication, though Meta has not publicly confirmed this and the write-up does not say whether data was accessed before patching.
What is Muse, and what can an AI agent do with a user's connected accounts?
Muse is Meta's personal AI agent, launched on September 8. It works through the online accounts a user connects to it. The article says it can make purchases on a user's behalf, so it can act across connected services rather than only answer questions in a chat.
Each person's agent runs on its own cloud computer. In the security demonstration, the researchers exploited a malicious page Muse opened and then chained browser and Linux kernel flaws to take control of that computer. The attack showed why the agent's access to accounts and private information matters: a weakness in its environment could expose more than the conversation itself.
Muse was designed on the assumption that its agent might be attacked, with limits intended to reduce potential damage. The article says passwords and tokens for connected accounts were supposed to remain beyond the agent's reach. NebuSec says the weaknesses it described have been fixed, but Meta has not publicly confirmed the fixes or said whether any user data was accessed before patching.
What was Sentinel meant to control, and why were connected-account passwords and tokens kept beyond the agent's reach?
Sentinel is a monitoring layer meant to be the single gate for anything Muse sends out. The design also keeps passwords and tokens for connected accounts beyond the agent's reach. Those controls matter because Muse can work through users' online accounts and make purchases on their behalf, giving the assistant access to consequential actions and private information.
The researchers' demonstration showed what happened when an attack got beyond the agent's normal boundaries. A malicious page could extract conversations and memories; chained browser and Linux kernel flaws then gave researchers full control of the user's cloud computer. That let them reach the credential store and Sentinel itself, which the design intended to wall off from the agent.
Muse's team described its approach as assuming the agent could be under attack and limiting potential damage. NebuSec said the controls worked as designed, yet the attack still ran through them. The firm says the weaknesses were fixed before publication. Meta has not publicly confirmed those fixes, and it is unknown from the write-up whether data was accessed before patching.
How could flaws in a browser and the Linux kernel combine to give a webpage control of a cloud computer?
A chain of flaws means one weakness can provide a way to reach another. Here, the researchers started with a malicious page opened by Muse. They then used a flaw in Muse's built-in browser together with a Linux kernel flaw. The article does not give the technical steps in detail, but says the combination let them take full control of the user's cloud computer.
That control mattered because the cloud computer held more than the agent's conversation history. It also contained saved files, the credential store, and Sentinel. The attack therefore crossed boundaries that were meant to keep sensitive components beyond the agent's reach. The page was not shown simply displaying malicious content; it was part of a route into the system running Muse.
The demonstration remained inside one user's cloud machine. The researchers did not break out to the shared server beneath it or other users' machines. NebuSec says the weaknesses it described were fixed before publication, while Meta has not publicly confirmed the fixes. The work demonstrates a possible chain, not a reported attack on real users.
What does it mean for each user's agent to run on its own cloud computer, and how can that limit the damage from a break-in?
The article says each person's Muse agent runs on its own cloud computer. In practice, that means the agent has a separate machine in the cloud for its work, while a shared server sits beneath those individual machines. The researchers took control of one user's computer, but did not break out to that shared server or other users' machines.
This separation can limit how far a break-in spreads. The malicious page and chained browser and kernel flaws compromised the affected user's cloud machine, including its files, credential store, and Sentinel. But the demonstration did not cross the boundary into the underlying shared server or another user's machine. It contained the observed reach of this particular attack.
That is a limit, not proof that every part of Muse's security design prevented compromise: the researchers still reached components meant to be walled off from the agent. The work was a laboratory demonstration, not an attack on real users. NebuSec says Muse's team fixed the reported weaknesses before publication, though Meta has not publicly confirmed the fixes.
Why can security defenses fail when attackers find a path through several weaknesses, even if each defense is designed to contain a breach?
Security defenses often protect different boundaries. If an attacker finds a weakness that crosses one boundary, it may provide a route to a component protected by another. A sequence of weaknesses can therefore have a larger effect than any single flaw. The article's demonstration is an example: the attack moved from a malicious page through browser and Linux kernel flaws to full control of the cloud machine.
Once in control, the researchers reached the credential store, saved files, and Sentinel. Those were parts of Muse's design meant to stay beyond the agent's reach. NebuSec said each control worked as designed, yet the attack still ran the whole way through. The report thus showed how defenses can operate as intended individually while a chain of vulnerabilities still reaches protected assets.
The compromise stopped at one user's cloud machine; it did not reach the shared server beneath it or other users' machines. NebuSec says the weaknesses were fixed before publication. Meta has not publicly confirmed the fixes, and the report does not say whether user data was accessed before patching.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web